“…Trace its technological development,at first, the dominant strategy was based on sentence templates and heuristics derived from empirical studies [6][7][8][9][10]. Starting around 2016, data-driven strategies based on neural networks came to the forefront, leveraging gains from both the AI/NLP and mining software repositories research communities [11][12][13][14].As far as we know, the existing deep learning based comment generation approaches mainly utilize the seq2seq model in which the program code is encoded into hidden space first and then decode it to produce the target comment.However, these kind of approaches have the following drawbacks: (1) they mainly take the source code as plain text and ignore the hierarchical structure of the source code; (2) most of the approaches only consider simple features,such as,tokens,which overlooking the hidden information that can help grab the relationships between source code and comments; (3) they typically train the decoder to produce the code annotation by calculating and maximizing the odds based on the subsequent natural language words, however in fact, they mainly produce the code annotation from scratch.…”